1 //===- ARMISelLowering.h - ARM DAG Lowering Interface -----------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines the interfaces that ARM uses to lower LLVM code into a 10 // selection DAG. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 15 #define LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 16 17 #include "MCTargetDesc/ARMBaseInfo.h" 18 #include "llvm/ADT/SmallVector.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/CodeGen/CallingConvLower.h" 21 #include "llvm/CodeGen/ISDOpcodes.h" 22 #include "llvm/CodeGen/MachineFunction.h" 23 #include "llvm/CodeGen/SelectionDAGNodes.h" 24 #include "llvm/CodeGen/TargetLowering.h" 25 #include "llvm/CodeGen/ValueTypes.h" 26 #include "llvm/IR/Attributes.h" 27 #include "llvm/IR/CallingConv.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/IRBuilder.h" 30 #include "llvm/IR/InlineAsm.h" 31 #include "llvm/Support/CodeGen.h" 32 #include "llvm/Support/MachineValueType.h" 33 #include <utility> 34 35 namespace llvm { 36 37 class ARMSubtarget; 38 class DataLayout; 39 class FastISel; 40 class FunctionLoweringInfo; 41 class GlobalValue; 42 class InstrItineraryData; 43 class Instruction; 44 class MachineBasicBlock; 45 class MachineInstr; 46 class SelectionDAG; 47 class TargetLibraryInfo; 48 class TargetMachine; 49 class TargetRegisterInfo; 50 class VectorType; 51 52 namespace ARMISD { 53 54 // ARM Specific DAG Nodes 55 enum NodeType : unsigned { 56 // Start the numbering where the builtin ops and target ops leave off. 57 FIRST_NUMBER = ISD::BUILTIN_OP_END, 58 59 Wrapper, // Wrapper - A wrapper node for TargetConstantPool, 60 // TargetExternalSymbol, and TargetGlobalAddress. 61 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in 62 // PIC mode. 63 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable 64 65 // Add pseudo op to model memcpy for struct byval. 66 COPY_STRUCT_BYVAL, 67 68 CALL, // Function call. 69 CALL_PRED, // Function call that's predicable. 70 CALL_NOLINK, // Function call with branch not branch-and-link. 71 tSECALL, // CMSE non-secure function call. 72 BRCOND, // Conditional branch. 73 BR_JT, // Jumptable branch. 74 BR2_JT, // Jumptable branch (2 level - jumptable entry is a jump). 75 RET_FLAG, // Return with a flag operand. 76 SERET_FLAG, // CMSE Entry function return with a flag operand. 77 INTRET_FLAG, // Interrupt return with an LR-offset and a flag operand. 78 79 PIC_ADD, // Add with a PC operand and a PIC label. 80 81 ASRL, // MVE long arithmetic shift right. 82 LSRL, // MVE long shift right. 83 LSLL, // MVE long shift left. 84 85 CMP, // ARM compare instructions. 86 CMN, // ARM CMN instructions. 87 CMPZ, // ARM compare that sets only Z flag. 88 CMPFP, // ARM VFP compare instruction, sets FPSCR. 89 CMPFPE, // ARM VFP signalling compare instruction, sets FPSCR. 90 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR. 91 CMPFPEw0, // ARM VFP signalling compare against zero instruction, sets 92 // FPSCR. 93 FMSTAT, // ARM fmstat instruction. 94 95 CMOV, // ARM conditional move instructions. 96 SUBS, // Flag-setting subtraction. 97 98 SSAT, // Signed saturation 99 USAT, // Unsigned saturation 100 101 BCC_i64, 102 103 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out. 104 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out. 105 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag. 106 107 ADDC, // Add with carry 108 ADDE, // Add using carry 109 SUBC, // Sub with carry 110 SUBE, // Sub using carry 111 LSLS, // Shift left producing carry 112 113 VMOVRRD, // double to two gprs. 114 VMOVDRR, // Two gprs to double. 115 VMOVSR, // move gpr to single, used for f32 literal constructed in a gpr 116 117 EH_SJLJ_SETJMP, // SjLj exception handling setjmp. 118 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp. 119 EH_SJLJ_SETUP_DISPATCH, // SjLj exception handling setup_dispatch. 120 121 TC_RETURN, // Tail call return pseudo. 122 123 THREAD_POINTER, 124 125 DYN_ALLOC, // Dynamic allocation on the stack. 126 127 MEMBARRIER_MCR, // Memory barrier (MCR) 128 129 PRELOAD, // Preload 130 131 WIN__CHKSTK, // Windows' __chkstk call to do stack probing. 132 WIN__DBZCHK, // Windows' divide by zero check 133 134 WLS, // Low-overhead loops, While Loop Start branch. See t2WhileLoopStart 135 WLSSETUP, // Setup for the iteration count of a WLS. See t2WhileLoopSetup. 136 LOOP_DEC, // Really a part of LE, performs the sub 137 LE, // Low-overhead loops, Loop End 138 139 PREDICATE_CAST, // Predicate cast for MVE i1 types 140 VECTOR_REG_CAST, // Reinterpret the current contents of a vector register 141 142 VCMP, // Vector compare. 143 VCMPZ, // Vector compare to zero. 144 VTST, // Vector test bits. 145 146 // Vector shift by vector 147 VSHLs, // ...left/right by signed 148 VSHLu, // ...left/right by unsigned 149 150 // Vector shift by immediate: 151 VSHLIMM, // ...left 152 VSHRsIMM, // ...right (signed) 153 VSHRuIMM, // ...right (unsigned) 154 155 // Vector rounding shift by immediate: 156 VRSHRsIMM, // ...right (signed) 157 VRSHRuIMM, // ...right (unsigned) 158 VRSHRNIMM, // ...right narrow 159 160 // Vector saturating shift by immediate: 161 VQSHLsIMM, // ...left (signed) 162 VQSHLuIMM, // ...left (unsigned) 163 VQSHLsuIMM, // ...left (signed to unsigned) 164 VQSHRNsIMM, // ...right narrow (signed) 165 VQSHRNuIMM, // ...right narrow (unsigned) 166 VQSHRNsuIMM, // ...right narrow (signed to unsigned) 167 168 // Vector saturating rounding shift by immediate: 169 VQRSHRNsIMM, // ...right narrow (signed) 170 VQRSHRNuIMM, // ...right narrow (unsigned) 171 VQRSHRNsuIMM, // ...right narrow (signed to unsigned) 172 173 // Vector shift and insert: 174 VSLIIMM, // ...left 175 VSRIIMM, // ...right 176 177 // Vector get lane (VMOV scalar to ARM core register) 178 // (These are used for 8- and 16-bit element types only.) 179 VGETLANEu, // zero-extend vector extract element 180 VGETLANEs, // sign-extend vector extract element 181 182 // Vector move immediate and move negated immediate: 183 VMOVIMM, 184 VMVNIMM, 185 186 // Vector move f32 immediate: 187 VMOVFPIMM, 188 189 // Move H <-> R, clearing top 16 bits 190 VMOVrh, 191 VMOVhr, 192 193 // Vector duplicate: 194 VDUP, 195 VDUPLANE, 196 197 // Vector shuffles: 198 VEXT, // extract 199 VREV64, // reverse elements within 64-bit doublewords 200 VREV32, // reverse elements within 32-bit words 201 VREV16, // reverse elements within 16-bit halfwords 202 VZIP, // zip (interleave) 203 VUZP, // unzip (deinterleave) 204 VTRN, // transpose 205 VTBL1, // 1-register shuffle with mask 206 VTBL2, // 2-register shuffle with mask 207 VMOVN, // MVE vmovn 208 209 // MVE Saturating truncates 210 VQMOVNs, // Vector (V) Saturating (Q) Move and Narrow (N), signed (s) 211 VQMOVNu, // Vector (V) Saturating (Q) Move and Narrow (N), unsigned (u) 212 213 // MVE float <> half converts 214 VCVTN, // MVE vcvt f32 -> f16, truncating into either the bottom or top 215 // lanes 216 VCVTL, // MVE vcvt f16 -> f32, extending from either the bottom or top lanes 217 218 // MVE VIDUP instruction, taking a start value and increment. 219 VIDUP, 220 221 // Vector multiply long: 222 VMULLs, // ...signed 223 VMULLu, // ...unsigned 224 225 VQDMULH, // MVE vqdmulh instruction 226 227 // MVE reductions 228 VADDVs, // sign- or zero-extend the elements of a vector to i32, 229 VADDVu, // add them all together, and return an i32 of their sum 230 VADDVps, // Same as VADDV[su] but with a v4i1 predicate mask 231 VADDVpu, 232 VADDLVs, // sign- or zero-extend elements to i64 and sum, returning 233 VADDLVu, // the low and high 32-bit halves of the sum 234 VADDLVAs, // Same as VADDLV[su] but also add an input accumulator 235 VADDLVAu, // provided as low and high halves 236 VADDLVps, // Same as VADDLV[su] but with a v4i1 predicate mask 237 VADDLVpu, 238 VADDLVAps, // Same as VADDLVp[su] but with a v4i1 predicate mask 239 VADDLVApu, 240 VMLAVs, // sign- or zero-extend the elements of two vectors to i32, multiply 241 // them 242 VMLAVu, // and add the results together, returning an i32 of their sum 243 VMLAVps, // Same as VMLAV[su] with a v4i1 predicate mask 244 VMLAVpu, 245 VMLALVs, // Same as VMLAV but with i64, returning the low and 246 VMLALVu, // high 32-bit halves of the sum 247 VMLALVps, // Same as VMLALV[su] with a v4i1 predicate mask 248 VMLALVpu, 249 VMLALVAs, // Same as VMLALV but also add an input accumulator 250 VMLALVAu, // provided as low and high halves 251 VMLALVAps, // Same as VMLALVA[su] with a v4i1 predicate mask 252 VMLALVApu, 253 VMINVu, // Find minimum unsigned value of a vector and register 254 VMINVs, // Find minimum signed value of a vector and register 255 VMAXVu, // Find maximum unsigned value of a vector and register 256 VMAXVs, // Find maximum signed value of a vector and register 257 258 SMULWB, // Signed multiply word by half word, bottom 259 SMULWT, // Signed multiply word by half word, top 260 UMLAL, // 64bit Unsigned Accumulate Multiply 261 SMLAL, // 64bit Signed Accumulate Multiply 262 UMAAL, // 64-bit Unsigned Accumulate Accumulate Multiply 263 SMLALBB, // 64-bit signed accumulate multiply bottom, bottom 16 264 SMLALBT, // 64-bit signed accumulate multiply bottom, top 16 265 SMLALTB, // 64-bit signed accumulate multiply top, bottom 16 266 SMLALTT, // 64-bit signed accumulate multiply top, top 16 267 SMLALD, // Signed multiply accumulate long dual 268 SMLALDX, // Signed multiply accumulate long dual exchange 269 SMLSLD, // Signed multiply subtract long dual 270 SMLSLDX, // Signed multiply subtract long dual exchange 271 SMMLAR, // Signed multiply long, round and add 272 SMMLSR, // Signed multiply long, subtract and round 273 274 // Single Lane QADD8 and QADD16. Only the bottom lane. That's what the b 275 // stands for. 276 QADD8b, 277 QSUB8b, 278 QADD16b, 279 QSUB16b, 280 281 // Operands of the standard BUILD_VECTOR node are not legalized, which 282 // is fine if BUILD_VECTORs are always lowered to shuffles or other 283 // operations, but for ARM some BUILD_VECTORs are legal as-is and their 284 // operands need to be legalized. Define an ARM-specific version of 285 // BUILD_VECTOR for this purpose. 286 BUILD_VECTOR, 287 288 // Bit-field insert 289 BFI, 290 291 // Vector OR with immediate 292 VORRIMM, 293 // Vector AND with NOT of immediate 294 VBICIMM, 295 296 // Pseudo vector bitwise select 297 VBSP, 298 299 // Pseudo-instruction representing a memory copy using ldm/stm 300 // instructions. 301 MEMCPY, 302 303 // Pseudo-instruction representing a memory copy using a tail predicated 304 // loop 305 MEMCPYLOOP, 306 // Pseudo-instruction representing a memset using a tail predicated 307 // loop 308 MEMSETLOOP, 309 310 // V8.1MMainline condition select 311 CSINV, // Conditional select invert. 312 CSNEG, // Conditional select negate. 313 CSINC, // Conditional select increment. 314 315 // Vector load N-element structure to all lanes: 316 VLD1DUP = ISD::FIRST_TARGET_MEMORY_OPCODE, 317 VLD2DUP, 318 VLD3DUP, 319 VLD4DUP, 320 321 // NEON loads with post-increment base updates: 322 VLD1_UPD, 323 VLD2_UPD, 324 VLD3_UPD, 325 VLD4_UPD, 326 VLD2LN_UPD, 327 VLD3LN_UPD, 328 VLD4LN_UPD, 329 VLD1DUP_UPD, 330 VLD2DUP_UPD, 331 VLD3DUP_UPD, 332 VLD4DUP_UPD, 333 334 // NEON stores with post-increment base updates: 335 VST1_UPD, 336 VST2_UPD, 337 VST3_UPD, 338 VST4_UPD, 339 VST2LN_UPD, 340 VST3LN_UPD, 341 VST4LN_UPD, 342 VST1x2_UPD, 343 VST1x3_UPD, 344 VST1x4_UPD, 345 346 // Load/Store of dual registers 347 LDRD, 348 STRD 349 }; 350 351 } // end namespace ARMISD 352 353 namespace ARM { 354 /// Possible values of current rounding mode, which is specified in bits 355 /// 23:22 of FPSCR. 356 enum Rounding { 357 RN = 0, // Round to Nearest 358 RP = 1, // Round towards Plus infinity 359 RM = 2, // Round towards Minus infinity 360 RZ = 3, // Round towards Zero 361 rmMask = 3 // Bit mask selecting rounding mode 362 }; 363 364 // Bit position of rounding mode bits in FPSCR. 365 const unsigned RoundingBitsPos = 22; 366 } // namespace ARM 367 368 /// Define some predicates that are used for node matching. 369 namespace ARM { 370 371 bool isBitFieldInvertedMask(unsigned v); 372 373 } // end namespace ARM 374 375 //===--------------------------------------------------------------------===// 376 // ARMTargetLowering - ARM Implementation of the TargetLowering interface 377 378 class ARMTargetLowering : public TargetLowering { 379 public: 380 explicit ARMTargetLowering(const TargetMachine &TM, 381 const ARMSubtarget &STI); 382 383 unsigned getJumpTableEncoding() const override; 384 bool useSoftFloat() const override; 385 386 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 387 388 /// ReplaceNodeResults - Replace the results of node with an illegal result 389 /// type with new values built out of custom code. 390 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 391 SelectionDAG &DAG) const override; 392 393 const char *getTargetNodeName(unsigned Opcode) const override; 394 isSelectSupported(SelectSupportKind Kind)395 bool isSelectSupported(SelectSupportKind Kind) const override { 396 // ARM does not support scalar condition selects on vectors. 397 return (Kind != ScalarCondVectorVal); 398 } 399 400 bool isReadOnly(const GlobalValue *GV) const; 401 402 /// getSetCCResultType - Return the value type to use for ISD::SETCC. 403 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, 404 EVT VT) const override; 405 406 MachineBasicBlock * 407 EmitInstrWithCustomInserter(MachineInstr &MI, 408 MachineBasicBlock *MBB) const override; 409 410 void AdjustInstrPostInstrSelection(MachineInstr &MI, 411 SDNode *Node) const override; 412 413 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const; 414 SDValue PerformBRCONDCombine(SDNode *N, SelectionDAG &DAG) const; 415 SDValue PerformCMOVToBFICombine(SDNode *N, SelectionDAG &DAG) const; 416 SDValue PerformIntrinsicCombine(SDNode *N, DAGCombinerInfo &DCI) const; 417 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 418 419 bool SimplifyDemandedBitsForTargetNode(SDValue Op, 420 const APInt &OriginalDemandedBits, 421 const APInt &OriginalDemandedElts, 422 KnownBits &Known, 423 TargetLoweringOpt &TLO, 424 unsigned Depth) const override; 425 426 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const override; 427 428 /// allowsMisalignedMemoryAccesses - Returns true if the target allows 429 /// unaligned memory accesses of the specified type. Returns whether it 430 /// is "fast" by reference in the second argument. 431 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AddrSpace, 432 Align Alignment, 433 MachineMemOperand::Flags Flags, 434 bool *Fast) const override; 435 436 EVT getOptimalMemOpType(const MemOp &Op, 437 const AttributeList &FuncAttributes) const override; 438 439 bool isTruncateFree(Type *SrcTy, Type *DstTy) const override; 440 bool isTruncateFree(EVT SrcVT, EVT DstVT) const override; 441 bool isZExtFree(SDValue Val, EVT VT2) const override; 442 bool shouldSinkOperands(Instruction *I, 443 SmallVectorImpl<Use *> &Ops) const override; 444 Type* shouldConvertSplatType(ShuffleVectorInst* SVI) const override; 445 446 bool isFNegFree(EVT VT) const override; 447 448 bool isVectorLoadExtDesirable(SDValue ExtVal) const override; 449 450 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override; 451 452 453 /// isLegalAddressingMode - Return true if the addressing mode represented 454 /// by AM is legal for this target, for a load/store of the specified type. 455 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, 456 Type *Ty, unsigned AS, 457 Instruction *I = nullptr) const override; 458 459 /// getScalingFactorCost - Return the cost of the scaling used in 460 /// addressing mode represented by AM. 461 /// If the AM is supported, the return value must be >= 0. 462 /// If the AM is not supported, the return value must be negative. 463 InstructionCost getScalingFactorCost(const DataLayout &DL, 464 const AddrMode &AM, Type *Ty, 465 unsigned AS) const override; 466 467 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const; 468 469 /// Returns true if the addressing mode representing by AM is legal 470 /// for the Thumb1 target, for a load/store of the specified type. 471 bool isLegalT1ScaledAddressingMode(const AddrMode &AM, EVT VT) const; 472 473 /// isLegalICmpImmediate - Return true if the specified immediate is legal 474 /// icmp immediate, that is the target has icmp instructions which can 475 /// compare a register against the immediate without having to materialize 476 /// the immediate into a register. 477 bool isLegalICmpImmediate(int64_t Imm) const override; 478 479 /// isLegalAddImmediate - Return true if the specified immediate is legal 480 /// add immediate, that is the target has add instructions which can 481 /// add a register and the immediate without having to materialize 482 /// the immediate into a register. 483 bool isLegalAddImmediate(int64_t Imm) const override; 484 485 /// getPreIndexedAddressParts - returns true by value, base pointer and 486 /// offset pointer and addressing mode by reference if the node's address 487 /// can be legally represented as pre-indexed load / store address. 488 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset, 489 ISD::MemIndexedMode &AM, 490 SelectionDAG &DAG) const override; 491 492 /// getPostIndexedAddressParts - returns true by value, base pointer and 493 /// offset pointer and addressing mode by reference if this node can be 494 /// combined with a load / store to form a post-indexed load / store. 495 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base, 496 SDValue &Offset, ISD::MemIndexedMode &AM, 497 SelectionDAG &DAG) const override; 498 499 void computeKnownBitsForTargetNode(const SDValue Op, KnownBits &Known, 500 const APInt &DemandedElts, 501 const SelectionDAG &DAG, 502 unsigned Depth) const override; 503 504 bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits, 505 const APInt &DemandedElts, 506 TargetLoweringOpt &TLO) const override; 507 508 bool ExpandInlineAsm(CallInst *CI) const override; 509 510 ConstraintType getConstraintType(StringRef Constraint) const override; 511 512 /// Examine constraint string and operand type and determine a weight value. 513 /// The operand object must already have been set up with the operand type. 514 ConstraintWeight getSingleConstraintMatchWeight( 515 AsmOperandInfo &info, const char *constraint) const override; 516 517 std::pair<unsigned, const TargetRegisterClass *> 518 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 519 StringRef Constraint, MVT VT) const override; 520 521 const char *LowerXConstraint(EVT ConstraintVT) const override; 522 523 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 524 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is 525 /// true it means one of the asm constraint of the inline asm instruction 526 /// being processed is 'm'. 527 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 528 std::vector<SDValue> &Ops, 529 SelectionDAG &DAG) const override; 530 531 unsigned getInlineAsmMemConstraint(StringRef ConstraintCode)532 getInlineAsmMemConstraint(StringRef ConstraintCode) const override { 533 if (ConstraintCode == "Q") 534 return InlineAsm::Constraint_Q; 535 else if (ConstraintCode.size() == 2) { 536 if (ConstraintCode[0] == 'U') { 537 switch(ConstraintCode[1]) { 538 default: 539 break; 540 case 'm': 541 return InlineAsm::Constraint_Um; 542 case 'n': 543 return InlineAsm::Constraint_Un; 544 case 'q': 545 return InlineAsm::Constraint_Uq; 546 case 's': 547 return InlineAsm::Constraint_Us; 548 case 't': 549 return InlineAsm::Constraint_Ut; 550 case 'v': 551 return InlineAsm::Constraint_Uv; 552 case 'y': 553 return InlineAsm::Constraint_Uy; 554 } 555 } 556 } 557 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); 558 } 559 getSubtarget()560 const ARMSubtarget* getSubtarget() const { 561 return Subtarget; 562 } 563 564 /// getRegClassFor - Return the register class that should be used for the 565 /// specified value type. 566 const TargetRegisterClass * 567 getRegClassFor(MVT VT, bool isDivergent = false) const override; 568 569 bool shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize, 570 unsigned &PrefAlign) const override; 571 572 /// createFastISel - This method returns a target specific FastISel object, 573 /// or null if the target does not support "fast" ISel. 574 FastISel *createFastISel(FunctionLoweringInfo &funcInfo, 575 const TargetLibraryInfo *libInfo) const override; 576 577 Sched::Preference getSchedulingPreference(SDNode *N) const override; 578 preferZeroCompareBranch()579 bool preferZeroCompareBranch() const override { return true; } 580 581 bool 582 isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override; 583 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override; 584 585 /// isFPImmLegal - Returns true if the target can instruction select the 586 /// specified FP immediate natively. If false, the legalizer will 587 /// materialize the FP immediate as a load from a constant pool. 588 bool isFPImmLegal(const APFloat &Imm, EVT VT, 589 bool ForCodeSize = false) const override; 590 591 bool getTgtMemIntrinsic(IntrinsicInfo &Info, 592 const CallInst &I, 593 MachineFunction &MF, 594 unsigned Intrinsic) const override; 595 596 /// Returns true if it is beneficial to convert a load of a constant 597 /// to just the constant itself. 598 bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 599 Type *Ty) const override; 600 601 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type 602 /// with this index. 603 bool isExtractSubvectorCheap(EVT ResVT, EVT SrcVT, 604 unsigned Index) const override; 605 shouldFormOverflowOp(unsigned Opcode,EVT VT,bool MathUsed)606 bool shouldFormOverflowOp(unsigned Opcode, EVT VT, 607 bool MathUsed) const override { 608 // Using overflow ops for overflow checks only should beneficial on ARM. 609 return TargetLowering::shouldFormOverflowOp(Opcode, VT, true); 610 } 611 612 /// Returns true if an argument of type Ty needs to be passed in a 613 /// contiguous block of registers in calling convention CallConv. 614 bool functionArgumentNeedsConsecutiveRegisters( 615 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override; 616 617 /// If a physical register, this returns the register that receives the 618 /// exception address on entry to an EH pad. 619 Register 620 getExceptionPointerRegister(const Constant *PersonalityFn) const override; 621 622 /// If a physical register, this returns the register that receives the 623 /// exception typeid on entry to a landing pad. 624 Register 625 getExceptionSelectorRegister(const Constant *PersonalityFn) const override; 626 627 Instruction *makeDMB(IRBuilder<> &Builder, ARM_MB::MemBOpt Domain) const; 628 Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr, 629 AtomicOrdering Ord) const override; 630 Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val, 631 Value *Addr, AtomicOrdering Ord) const override; 632 633 void emitAtomicCmpXchgNoStoreLLBalance(IRBuilder<> &Builder) const override; 634 635 Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst, 636 AtomicOrdering Ord) const override; 637 Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst, 638 AtomicOrdering Ord) const override; 639 640 unsigned getMaxSupportedInterleaveFactor() const override; 641 642 bool lowerInterleavedLoad(LoadInst *LI, 643 ArrayRef<ShuffleVectorInst *> Shuffles, 644 ArrayRef<unsigned> Indices, 645 unsigned Factor) const override; 646 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI, 647 unsigned Factor) const override; 648 649 bool shouldInsertFencesForAtomic(const Instruction *I) const override; 650 TargetLoweringBase::AtomicExpansionKind 651 shouldExpandAtomicLoadInIR(LoadInst *LI) const override; 652 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override; 653 TargetLoweringBase::AtomicExpansionKind 654 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override; 655 TargetLoweringBase::AtomicExpansionKind 656 shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override; 657 658 bool useLoadStackGuardNode() const override; 659 660 void insertSSPDeclarations(Module &M) const override; 661 Value *getSDagStackGuard(const Module &M) const override; 662 Function *getSSPStackGuardCheck(const Module &M) const override; 663 664 bool canCombineStoreAndExtract(Type *VectorTy, Value *Idx, 665 unsigned &Cost) const override; 666 canMergeStoresTo(unsigned AddressSpace,EVT MemVT,const SelectionDAG & DAG)667 bool canMergeStoresTo(unsigned AddressSpace, EVT MemVT, 668 const SelectionDAG &DAG) const override { 669 // Do not merge to larger than i32. 670 return (MemVT.getSizeInBits() <= 32); 671 } 672 673 bool isCheapToSpeculateCttz() const override; 674 bool isCheapToSpeculateCtlz() const override; 675 convertSetCCLogicToBitwiseLogic(EVT VT)676 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override { 677 return VT.isScalarInteger(); 678 } 679 supportSwiftError()680 bool supportSwiftError() const override { 681 return true; 682 } 683 hasStandaloneRem(EVT VT)684 bool hasStandaloneRem(EVT VT) const override { 685 return HasStandaloneRem; 686 } 687 688 bool shouldExpandShift(SelectionDAG &DAG, SDNode *N) const override; 689 690 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool isVarArg) const; 691 CCAssignFn *CCAssignFnForReturn(CallingConv::ID CC, bool isVarArg) const; 692 693 /// Returns true if \p VecTy is a legal interleaved access type. This 694 /// function checks the vector element type and the overall width of the 695 /// vector. 696 bool isLegalInterleavedAccessType(unsigned Factor, FixedVectorType *VecTy, 697 Align Alignment, 698 const DataLayout &DL) const; 699 700 bool alignLoopsWithOptSize() const override; 701 702 /// Returns the number of interleaved accesses that will be generated when 703 /// lowering accesses of the given type. 704 unsigned getNumInterleavedAccesses(VectorType *VecTy, 705 const DataLayout &DL) const; 706 707 void finalizeLowering(MachineFunction &MF) const override; 708 709 /// Return the correct alignment for the current calling convention. 710 Align getABIAlignmentForCallingConv(Type *ArgTy, 711 DataLayout DL) const override; 712 713 bool isDesirableToCommuteWithShift(const SDNode *N, 714 CombineLevel Level) const override; 715 716 bool shouldFoldConstantShiftPairToMask(const SDNode *N, 717 CombineLevel Level) const override; 718 719 bool preferIncOfAddToSubOfNot(EVT VT) const override; 720 721 protected: 722 std::pair<const TargetRegisterClass *, uint8_t> 723 findRepresentativeClass(const TargetRegisterInfo *TRI, 724 MVT VT) const override; 725 726 private: 727 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can 728 /// make the right decision when generating code for different targets. 729 const ARMSubtarget *Subtarget; 730 731 const TargetRegisterInfo *RegInfo; 732 733 const InstrItineraryData *Itins; 734 735 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created. 736 unsigned ARMPCLabelIndex; 737 738 // TODO: remove this, and have shouldInsertFencesForAtomic do the proper 739 // check. 740 bool InsertFencesForAtomic; 741 742 bool HasStandaloneRem = true; 743 744 void addTypeForNEON(MVT VT, MVT PromotedLdStVT, MVT PromotedBitwiseVT); 745 void addDRTypeForNEON(MVT VT); 746 void addQRTypeForNEON(MVT VT); 747 std::pair<SDValue, SDValue> getARMXALUOOp(SDValue Op, SelectionDAG &DAG, SDValue &ARMcc) const; 748 749 using RegsToPassVector = SmallVector<std::pair<unsigned, SDValue>, 8>; 750 751 void PassF64ArgInRegs(const SDLoc &dl, SelectionDAG &DAG, SDValue Chain, 752 SDValue &Arg, RegsToPassVector &RegsToPass, 753 CCValAssign &VA, CCValAssign &NextVA, 754 SDValue &StackPtr, 755 SmallVectorImpl<SDValue> &MemOpChains, 756 ISD::ArgFlagsTy Flags) const; 757 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 758 SDValue &Root, SelectionDAG &DAG, 759 const SDLoc &dl) const; 760 761 CallingConv::ID getEffectiveCallingConv(CallingConv::ID CC, 762 bool isVarArg) const; 763 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return, 764 bool isVarArg) const; 765 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg, 766 const SDLoc &dl, SelectionDAG &DAG, 767 const CCValAssign &VA, 768 ISD::ArgFlagsTy Flags) const; 769 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const; 770 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const; 771 SDValue LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const; 772 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG, 773 const ARMSubtarget *Subtarget) const; 774 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 775 const ARMSubtarget *Subtarget) const; 776 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 777 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; 778 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; 779 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const; 780 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const; 781 SDValue LowerGlobalAddressWindows(SDValue Op, SelectionDAG &DAG) const; 782 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; 783 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 784 SelectionDAG &DAG) const; 785 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA, 786 SelectionDAG &DAG, 787 TLSModel::Model model) const; 788 SDValue LowerGlobalTLSAddressDarwin(SDValue Op, SelectionDAG &DAG) const; 789 SDValue LowerGlobalTLSAddressWindows(SDValue Op, SelectionDAG &DAG) const; 790 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const; 791 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const; 792 SDValue LowerSignedALUO(SDValue Op, SelectionDAG &DAG) const; 793 SDValue LowerUnsignedALUO(SDValue Op, SelectionDAG &DAG) const; 794 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const; 795 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; 796 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const; 797 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const; 798 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const; 799 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 800 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 801 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const; 802 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const; 803 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const; 804 SDValue LowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const; 805 SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG, 806 const ARMSubtarget *ST) const; 807 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 808 const ARMSubtarget *ST) const; 809 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 810 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const; 811 SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const; 812 SDValue LowerDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed) const; 813 void ExpandDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed, 814 SmallVectorImpl<SDValue> &Results) const; 815 SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG, 816 const ARMSubtarget *Subtarget) const; 817 SDValue LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG, bool Signed, 818 SDValue &Chain) const; 819 SDValue LowerREM(SDNode *N, SelectionDAG &DAG) const; 820 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; 821 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const; 822 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const; 823 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const; 824 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 825 SDValue LowerFSETCC(SDValue Op, SelectionDAG &DAG) const; 826 void lowerABS(SDNode *N, SmallVectorImpl<SDValue> &Results, 827 SelectionDAG &DAG) const; 828 void LowerLOAD(SDNode *N, SmallVectorImpl<SDValue> &Results, 829 SelectionDAG &DAG) const; 830 831 Register getRegisterByName(const char* RegName, LLT VT, 832 const MachineFunction &MF) const override; 833 834 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 835 SmallVectorImpl<SDNode *> &Created) const override; 836 837 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, 838 EVT VT) const override; 839 840 SDValue MoveToHPR(const SDLoc &dl, SelectionDAG &DAG, MVT LocVT, MVT ValVT, 841 SDValue Val) const; 842 SDValue MoveFromHPR(const SDLoc &dl, SelectionDAG &DAG, MVT LocVT, 843 MVT ValVT, SDValue Val) const; 844 845 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const; 846 847 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 848 CallingConv::ID CallConv, bool isVarArg, 849 const SmallVectorImpl<ISD::InputArg> &Ins, 850 const SDLoc &dl, SelectionDAG &DAG, 851 SmallVectorImpl<SDValue> &InVals, bool isThisReturn, 852 SDValue ThisVal) const; 853 supportSplitCSR(MachineFunction * MF)854 bool supportSplitCSR(MachineFunction *MF) const override { 855 return MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS && 856 MF->getFunction().hasFnAttribute(Attribute::NoUnwind); 857 } 858 859 void initializeSplitCSR(MachineBasicBlock *Entry) const override; 860 void insertCopiesSplitCSR( 861 MachineBasicBlock *Entry, 862 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override; 863 864 bool 865 splitValueIntoRegisterParts(SelectionDAG &DAG, const SDLoc &DL, SDValue Val, 866 SDValue *Parts, unsigned NumParts, MVT PartVT, 867 Optional<CallingConv::ID> CC) const override; 868 869 SDValue 870 joinRegisterPartsIntoValue(SelectionDAG &DAG, const SDLoc &DL, 871 const SDValue *Parts, unsigned NumParts, 872 MVT PartVT, EVT ValueVT, 873 Optional<CallingConv::ID> CC) const override; 874 875 SDValue 876 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, 877 const SmallVectorImpl<ISD::InputArg> &Ins, 878 const SDLoc &dl, SelectionDAG &DAG, 879 SmallVectorImpl<SDValue> &InVals) const override; 880 881 int StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG, const SDLoc &dl, 882 SDValue &Chain, const Value *OrigArg, 883 unsigned InRegsParamRecordIdx, int ArgOffset, 884 unsigned ArgSize) const; 885 886 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 887 const SDLoc &dl, SDValue &Chain, 888 unsigned ArgOffset, unsigned TotalArgRegsSaveSize, 889 bool ForceMutable = false) const; 890 891 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI, 892 SmallVectorImpl<SDValue> &InVals) const override; 893 894 /// HandleByVal - Target-specific cleanup for ByVal support. 895 void HandleByVal(CCState *, unsigned &, Align) const override; 896 897 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 898 /// for tail call optimization. Targets which want to do tail call 899 /// optimization should implement this function. 900 bool IsEligibleForTailCallOptimization( 901 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg, 902 bool isCalleeStructRet, bool isCallerStructRet, 903 const SmallVectorImpl<ISD::OutputArg> &Outs, 904 const SmallVectorImpl<SDValue> &OutVals, 905 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG, 906 const bool isIndirect) const; 907 908 bool CanLowerReturn(CallingConv::ID CallConv, 909 MachineFunction &MF, bool isVarArg, 910 const SmallVectorImpl<ISD::OutputArg> &Outs, 911 LLVMContext &Context) const override; 912 913 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, 914 const SmallVectorImpl<ISD::OutputArg> &Outs, 915 const SmallVectorImpl<SDValue> &OutVals, 916 const SDLoc &dl, SelectionDAG &DAG) const override; 917 918 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override; 919 920 bool mayBeEmittedAsTailCall(const CallInst *CI) const override; 921 shouldConsiderGEPOffsetSplit()922 bool shouldConsiderGEPOffsetSplit() const override { return true; } 923 924 bool isUnsupportedFloatingType(EVT VT) const; 925 926 SDValue getCMOV(const SDLoc &dl, EVT VT, SDValue FalseVal, SDValue TrueVal, 927 SDValue ARMcc, SDValue CCR, SDValue Cmp, 928 SelectionDAG &DAG) const; 929 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 930 SDValue &ARMcc, SelectionDAG &DAG, const SDLoc &dl) const; 931 SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 932 const SDLoc &dl, bool Signaling = false) const; 933 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const; 934 935 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const; 936 937 void SetupEntryBlockForSjLj(MachineInstr &MI, MachineBasicBlock *MBB, 938 MachineBasicBlock *DispatchBB, int FI) const; 939 940 void EmitSjLjDispatchBlock(MachineInstr &MI, MachineBasicBlock *MBB) const; 941 942 bool RemapAddSubWithFlags(MachineInstr &MI, MachineBasicBlock *BB) const; 943 944 MachineBasicBlock *EmitStructByval(MachineInstr &MI, 945 MachineBasicBlock *MBB) const; 946 947 MachineBasicBlock *EmitLowered__chkstk(MachineInstr &MI, 948 MachineBasicBlock *MBB) const; 949 MachineBasicBlock *EmitLowered__dbzchk(MachineInstr &MI, 950 MachineBasicBlock *MBB) const; 951 void addMVEVectorTypes(bool HasMVEFP); 952 void addAllExtLoads(const MVT From, const MVT To, LegalizeAction Action); 953 void setAllExpand(MVT VT); 954 }; 955 956 enum VMOVModImmType { 957 VMOVModImm, 958 VMVNModImm, 959 MVEVMVNModImm, 960 OtherModImm 961 }; 962 963 namespace ARM { 964 965 FastISel *createFastISel(FunctionLoweringInfo &funcInfo, 966 const TargetLibraryInfo *libInfo); 967 968 } // end namespace ARM 969 970 } // end namespace llvm 971 972 #endif // LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 973